Is Your Text AI-Generated?

What signals do AI detectors actually measure?

Modern AI detectors don't look at your words directly. They measure statistical patterns that human and AI writing produce in different distributions. The two foundational signals are perplexity and burstiness.

Perplexity measures how predictable each word is given everything that came before it. AI models like ChatGPT and Claude are trained to pick the statistically safest next word at every step, so their output has unusually low perplexity. Humans, even careful writers, occasionally pick a less-expected word, a metaphor, an idiom, or a regional turn of phrase. That bumps perplexity up.

Burstiness measures how much sentence length varies across a paragraph. Human writing has rhythm: a long sentence followed by a short one, then a fragment, then another long one. AI tends to produce a flat, uniform cadence. Sentences hover around 18 to 25 words for paragraphs at a time. Burstiness catches this.

Why GPTZero, Turnitin, and Copyleaks sometimes disagree

Different detectors weight different signals differently. GPTZero leans heavily on perplexity and burstiness. Turnitin combines those with a proprietary trained classifier that's tuned for academic writing specifically. Copyleaks adds plagiarism cross-referencing on top, which can flag legitimately-cited human writing as suspicious.

The result: the same text can score 95% AI on Turnitin and 30% AI on GPTZero. Neither is wrong. They're just measuring slightly different things. The smart workflow is to check against multiple detectors and humanize until you pass the strictest one your audience uses.

What our 8-signal detector adds beyond perplexity

WriteHumanly's detector goes beyond the two-signal academic baseline. We layer six additional signals to cut false positives on heavily-edited human writing and false negatives on lightly-paraphrased AI output:

• Vocabulary consistency: how tightly the writer sticks to one register across the document.
• Word repetition windows: AI reuses content words within 50-word spans more than humans do.
• Lexical diversity: the type-token ratio that distinguishes formulaic prose from natural variation.
• Formality fingerprint: counts of formal connectors, hedging language, and zero-contraction patterns.
• Stylometric profile: sentence-shape signatures unique to GPT, Claude, and Gemini training data.
• ChatGPT phrase density: weighted scoring for known AI clichés like “in conclusion,” “it's important to note,” and “delve into.”

How to read the score correctly

Our score is the probability that the text was written by AI, expressed as a percentage. Below 30% means the text reads as human across all eight signals. 30 to 55% means mixed signals. This often happens with AI-assisted writing where the human edited heavily but kept some AI structures. Above 55% indicates clear AI generation.

A 100% score doesn't mean we “know” the text was AI. It means our signals all align with AI-typical patterns to a degree that would be improbable for human writing. There's always uncertainty. Use the sentence-level breakdown to see exactly which sections are driving the score and decide for yourself.

What if you get a false positive?

Heavily-edited writing, technical documentation, formulaic genres (legal, medical, academic abstracts), and second-language writers can all trigger false AI flags. If your text is genuinely human but scoring high, the fix is the same as for actual AI text: run it through our humanizer. The humanizer doesn't make text “less original”. It injects natural variation that AI detectors don't flag, which is what your writing was missing structurally to begin with.